Neural networks robot controller trained with evolution strategies

Abstract

Neural networks (NN) can be used as controllers in autonomous robots. The specific features of the navigation problem in robotics make generation of good training sets for the NN difficult. An evolution strategy (ES) is introduced to learn the weights of the NN instead of the learning method of the network. The ES is used to learn high performance reactive behavior for navigation and collision avoidance. No subjective information about "how to accomplish the task" has been included in the fitness function. The learned behaviors are able to solve the problem in different environments; therefore, the learning process has the proven ability to obtain a specialized behavior. All the behaviors obtained have been tested in a set of environments and the capability of generalization is shown for each learned behavior. A simulator based on the mini-robot, Khepera, has been used to learn each behavior.